[Hackathon 6th Code Camp No.15] support neuraloperator #867

Yang-Changhui · 2024-04-25T02:52:15Z

PR types

New features

PR changes

Others

Describe

增加neuraloperator模型

paddle-bot · 2024-04-25T02:52:20Z

Thanks for your contribution!

zhiminzhang0830 · 2024-04-26T07:31:52Z

ppsci/arch/fno_block.py

+        except KeyError:
+            padding = [round(p * r) for (p, r) in zip(self.domain_padding, resolution)]
+
+            print(


建议使用ppsci内置的log

zhiminzhang0830 · 2024-04-26T07:32:56Z

ppsci/arch/fno_block.py

+            # of the "padding" list i.e. the last axis of tensor 'x' will be
+            # padded by the amount mention at the first position of the
+            # 'padding' vector. The details about F.pad can be found here:
+            # https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html


确认注释正确性

zhiminzhang0830 · 2024-04-26T07:40:36Z

ppsci/arch/fno_block.py

+            padded = F.pad(x, padding, mode="constant")
+
+            output_shape = padded.shape[2:]
+


删除多余空格

zhiminzhang0830 · 2024-04-26T07:44:11Z

ppsci/arch/fno_block.py

+        return nn.Identity()
+    else:
+        raise ValueError(
+            f"Got skip-connection type = {type}, expected one of {'soft-gating', 'linear', 'id'}."


id --> identity

zhiminzhang0830 · 2024-04-26T08:15:25Z

ppsci/arch/fno_block.py

+            arg  is only checked when `implementation=reconstructed`. Defaults to False.
+
+    Raises:
+        ValueError: _description_


删除raises

zhiminzhang0830 · 2024-04-26T08:15:57Z

ppsci/arch/fno_block.py

+
+    if implementation == "reconstructed":
+        if separable:
+            print("SEPARABLE")


删除print

zhiminzhang0830 · 2024-04-26T08:16:30Z

ppsci/arch/fno_block.py

+        #     elif weight.name.lower() == 'complexcp':
+        #         return _contract_cp
+        #     else:
+        #         raise ValueError(f'Got unexpected factorized weight type {weight.name}')


删除无意义的注释

zhiminzhang0830 · 2024-04-28T12:23:21Z

ppsci/arch/fno_block.py

+        fixed_rank_modes=False,
+        joint_factorization=False,
+        init_std="auto",
+        fft_norm="backward",


初始化添加类型提示，其他的类也注意一下

zhiminzhang0830 · 2024-04-28T12:27:34Z

ppsci/arch/fno_block.py

+            )
+
+    def forward(
+        self, x: paddle.Tensor, indices=0, output_shape: Optional[Tuple[int]] = None


变量 indices 添加类型提示

zhiminzhang0830 · 2024-04-28T12:29:20Z

ppsci/arch/fno_block.py

+
+    This is provided for reference only,
+    see :class:`neuralop.layers.SpectraConv` for the preferred, general implementation
+    """


更正注释，ppsci没有neuralop.layers.SpectraConv这个路径，另外确认该类是否在代码中有用到

这几个类都没有用到，但是是可选的

…/PaddleScience into add_neuraloperator

zhiminzhang0830 · 2024-04-29T04:05:00Z

ppsci/arch/sfnonet.py

+
+
+def _contract_dense_trick(x, weight_real, weight_imag, separable=False, dhconv=True):
+    # the same as above function, but do the complex multiplication manually to avoid the einsum bug in paddle


“the same as above function” ？这个函数上面没有其他函数呀？

zhiminzhang0830 · 2024-04-29T04:07:59Z

ppsci/arch/sfnonet.py

+
+    if implementation == "reconstructed":
+        if separable:
+            print("SEPARABLE")


确认是否有必要print

zhiminzhang0830 · 2024-04-29T04:08:24Z

ppsci/arch/sfnonet.py

+    elif implementation == "factorized":
+        if isinstance(weight, paddle.Tensor):
+            return _contract_dense_trick
+        # TODO: FactorizedTensor not supported yet


这个注释可以删掉吧

zhiminzhang0830 · 2024-04-29T04:15:25Z

ppsci/arch/sfnonet.py

+
+
+class SphericalConv(nn.Layer):
+    """Spherical Convolution, base class for the SFNO [1]_


注释最后用“.“更合适吧？

zhiminzhang0830 · 2024-04-29T04:16:30Z

ppsci/arch/sfnonet.py

+    """Spherical Convolution, base class for the SFNO [1].
+        .. [1] Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere,
+           Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, Anima Anandkumar,
+           ICML 2023.


references重复了

zhiminzhang0830 · 2024-04-29T04:17:16Z

ppsci/arch/sfnonet.py

+    Args:
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        n_modes (int): Number of modes to use for contraction in Fourier domain during


注释中的变量类型与初始化函数中的保持一致

zhiminzhang0830 · 2024-05-06T12:09:09Z

ppsci/arch/fno_block.py

+    """Applies domain padding scaled automatically to the input's resolution
+
+    Args:
+        domain_padding (float): typically, between zero and one, percentage of padding to use.


类型与初始化代码保持一致， Union[float, List[float]]

zhiminzhang0830 · 2024-05-10T09:27:10Z

ppsci/data/dataset/darcyflow_dataset.py

+
+        Returns
+        -------
+        torch.tensor


函数注释修改下，建议注释风格修改为PaddleScience风格

zhiminzhang0830 · 2024-05-10T09:27:44Z

ppsci/data/dataset/darcyflow_dataset.py

+
+        # x is in shape of batch*n or T*batch*n
+        # x = (x.view(self.sample_shape) * std) + mean
+        # x = x.view(-1, *self.sample_shape)


不需要的注释删除吧

zhiminzhang0830 · 2024-05-10T09:30:49Z

ppsci/data/dataset/darcyflow_dataset.py

+        return x
+
+
+def count_params(model):


确认此函数是否有使用

zhiminzhang0830 · 2024-05-10T09:40:55Z

ppsci/data/dataset/darcyflow_dataset.py

+    50 samples at resolution 32x32.
+
+    Args:
+        test_resolutions (List[int,...]): The resolutions to test dataset. Default is [16, 32].


完善docstring，与初始化函数的入参保持一致，缺少train_resolution，记得补充完整。test_resolutions参数的docstring的默认值也需要与初始化入参保持一致

zhiminzhang0830 · 2024-05-10T09:56:25Z

ppsci/data/dataset/spherical_swe_dataset.py

+    50 samples at resolution 64x128.
+
+    Args:
+        test_resolutions (List[str,...]): The resolutions to test dataset. Default is ["34x64","64x128"].


完善docstring

… into add_neuraloperator

…/PaddleScience into add_neuraloperator

zhiminzhang0830 · 2024-05-11T09:44:41Z

docs/zh/api/data/dataset.md

添加 SphericalSWEDataset

因为这个数据集是我使用pytorch原代码生成的数据，直接保存的npy文件，这样还需要写进去吗

zhiminzhang0830 · 2024-05-11T09:45:57Z

examples/neuraloperator/conf/sfno_swe_pretrain.yaml

+  pretrained_model_path: ./outputs_sfno_pretrain/checkpoints/best_model.pdparams
+  export_path: ./inference/uno/uno_darcyflow
+  pdmodel_path: ${INFER.export_path}.pdmodel
+  pdpiparams_path: ${INFER.export_path}.pdiparams


pdpiparams_path --> pdiparams_path 全部替换一下

zhiminzhang0830 · 2024-05-11T09:52:37Z

ppsci/data/dataset/darcyflow_dataset.py

+        encoding (str): The type of encoding. Default is 'channel-wise'.
+        channel_dim (int): The location of unsqueeze. Default is 1.
+            where to put the channel dimension, defaults size is batch, channel, height, width
+        training (str): Wether to use training or test dataset. Default is 'train'.


其他数据集中使用 training 参数表示该数据集是否用于训练，使用bool类型。在此数据集中，training 参数还用于表示测试时的数据选择，与变量名称”training“不太相符，建议修改为其他更有意义的变量名称

zhiminzhang0830 · 2024-05-11T09:53:16Z

ppsci/data/dataset/spherical_swe_dataset.py

+            Defaults to None.
+        test_resolutions (Tuple[str, ...], optional): The resolutions to test dataset. Defaults to ["34x64", "64x128"].
+        train_resolution (str, optional): The resolutions to train dataset. Defaults to "34x64".
+        training (str, optional): Wether to use training or test dataset. Defaults to "train".


zhiminzhang0830 · 2024-05-11T09:54:15Z

examples/neuraloperator/earthformer.md

与本PR无关，记得删除

zhiminzhang0830 · 2024-05-11T09:58:19Z

examples/neuraloperator/train_sfno.py

+    input_data = data["x"][0].reshape(1, *data["x"].shape[1:]).astype("float32")
+    label = data["y"][0][0, ...].astype("float32")
+
+    model = ppsci.arch.SFNONet(


好像还是使用动态图的形式进行预测的？并没有使用export的静态图模型进行预测？

是动态图形式进行预测的，因为导出静态图模型进行预测时，结果和动态图不一样，可能是因为使用虚数的缘故，也可能是独有的API导致的导出模型的网络权重有问题，跟动态图的网络权重不一样，这个之前给你说过

…/PaddleScience into add_neuraloperator

Yang-Changhui · 2024-05-20T08:11:28Z

paddle使用dev版本

zhiminzhang0830 · 2024-05-26T05:20:09Z

examples/neuraloperator/train_uno.py

+        sup_validator_32.name: sup_validator_32,
+    }
+
+    model = ppsci.arch.TFNO2dNet(


这个地方应该是UNO的模型吧

zhiminzhang0830 · 2024-05-26T05:20:45Z

examples/neuraloperator/conf/tfno_darcyflow_pretrain.yaml

+    scheduler_T_max: 500
+  wd: 1.0e-4
+  batch_size: 16
+  pretrained_model_path: ./pretrainmodel/tfno_pretrain.pdparams


直接设置成null吧，另外两个也改下

zhiminzhang0830 · 2024-05-26T05:23:15Z

ppsci/arch/fno_block.py

+        n_layers: int = 1,
+        max_n_modes: int = None,
+        use_mlp: bool = False,
+        mlp: Optional[dict[float, float]] = None,


此处dict应该为Dict, 并在文件开头添加：from typing import Dict，其他文件也有类似问题，建议统一检查

zhiminzhang0830 · 2024-05-26T05:23:54Z

examples/neuraloperator/train_uno.py

+    ax.set_title("Model prediction")
+    plt.xticks([], [])
+    plt.yticks([], [])
+    plt.savefig(cfg.output_dir + "123.png")


此处123.png，作为文件名称有什么含义么

…/PaddleScience into add_neuraloperator

zhiminzhang0830 · 2024-05-26T10:09:44Z

examples/neuraloperator/train_tfno.py

+    solver.export(input_spec, cfg.INFER.export_path)
+
+
+# 使用静态图推理出错


现在推理还有问题么

zhiminzhang0830 · 2024-05-26T10:10:09Z

examples/neuraloperator/predictor.py

+        """
+        if batch_size != 1:
+            raise ValueError(
+                f"FNOPredictor only support batch_size=1, but got {batch_size}"


FNOPredictor --> SFNOPredictor

…/PaddleScience into add_neuraloperator

zhiminzhang0830

LGTM， TODO：add doc

) * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator * add-neuraloperator

add-neuraloperator

273d4b2

paddle-bot bot added the contributor label Apr 25, 2024

Yang-Changhui and others added 2 commits April 25, 2024 11:05

add-neuraloperator

f0b900e

Merge branch 'develop' into add_neuraloperator

f0760a7

zhiminzhang0830 reviewed Apr 28, 2024

View reviewed changes

Yang-Changhui and others added 7 commits April 29, 2024 11:47

add-neuraloperator

5c59594

add-neuralopertor

55e5b14

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

34009e8

…/PaddleScience into add_neuraloperator

add-neuraloperator

0a1cde8

add-neuraloperator

99fbd35

add-neuraloperator

fd8c49f

Merge branch 'develop' into add_neuraloperator

92db31c

zhiminzhang0830 reviewed May 10, 2024

View reviewed changes

Yang-Changhui and others added 7 commits May 10, 2024 20:05

Merge branch 'develop' into add_neuraloperator

1be2a63

add-neuraloperator

a8c5526

add-neuraloperator

11c50fd

Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleScience…

9dc7917

… into add_neuraloperator

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

a48e8fb

…/PaddleScience into add_neuraloperator

add-neuraloperator

b1df8c4

Merge branch 'develop' into add_neuraloperator

df747a9

Yang-Changhui mentioned this pull request May 10, 2024

[WeeklyReports] 2024.04.19~2024.05.10 周报汇总 PFCCLab/Camp#252

Closed

29 tasks

zhiminzhang0830 reviewed May 11, 2024

View reviewed changes

Yang-Changhui and others added 6 commits May 13, 2024 11:48

add-neuraloperator

2a4d0da

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

c91d92f

…/PaddleScience into add_neuraloperator

Merge branch 'develop' into add_neuraloperator

5e5ad68

add-neuraloperator

750bb7d

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

cb76494

…/PaddleScience into add_neuraloperator

Merge branch 'develop' into add_neuraloperator

9dcd49a

Yang-Changhui mentioned this pull request May 25, 2024

[WeeklyReports] 2024.05.11~2024.05.24 周报汇总 PFCCLab/Camp#280

Closed

29 tasks

zhiminzhang0830 reviewed May 26, 2024

View reviewed changes

Yang-Changhui and others added 3 commits May 26, 2024 14:47

add-neuraloperator

560592d

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

f1b197f

…/PaddleScience into add_neuraloperator

Merge branch 'develop' into add_neuraloperator

60aa2c8

zhiminzhang0830 reviewed May 26, 2024

View reviewed changes

Yang-Changhui added 2 commits May 26, 2024 19:24

add-neuraloperator

8d72bf7

Merge branch 'add_neuraloperator' of https://github.com/Yang-Changhui…

7b819fc

…/PaddleScience into add_neuraloperator

zhiminzhang0830 approved these changes May 27, 2024

View reviewed changes

zhiminzhang0830 merged commit aa963b8 into PaddlePaddle:develop May 27, 2024
3 of 4 checks passed

Yang-Changhui mentioned this pull request Jun 11, 2024

WAVE SUMMIT+2024上半年飞桨开源之星评选-信息征集 PaddlePaddle/community#892

Closed

		padded = F.pad(x, padding, mode="constant")

		output_shape = padded.shape[2:]



		def _contract_dense_trick(x, weight_real, weight_imag, separable=False, dhconv=True):
		# the same as above function, but do the complex multiplication manually to avoid the einsum bug in paddle



		class SphericalConv(nn.Layer):
		"""Spherical Convolution, base class for the SFNO [1]_

		solver.export(input_spec, cfg.INFER.export_path)


		# 使用静态图推理出错

[Hackathon 6th Code Camp No.15] support neuraloperator #867

[Hackathon 6th Code Camp No.15] support neuraloperator #867

Conversation

Yang-Changhui commented Apr 25, 2024

PR types

PR changes

Describe

paddle-bot bot commented Apr 25, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zhiminzhang0830 May 10, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Yang-Changhui commented May 20, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zhiminzhang0830 left a comment

Choose a reason for hiding this comment

zhiminzhang0830 May 10, 2024 •

edited

Loading